Assessing Impact of Data Partitioning for Approximate Memory in C/C++ Code

Abstract

Approximate memory is a technique to mitigate the performance gap between memory subsystems and CPUs with its reduced access latency at a cost of data integrity. To gain benefit from approximate memory for realistic applications, it is crucial to partition applications' data to approximate data and critical data and apply different error rates. However, error rates cannot be controlled in a fine-grained manner (e.g., per byte) due to fundamental limitations of how approximate memory can be realized. Due to this, if approximate data and critical data are interleaved in a data structure (e.g., a C struct that has a pointer and an approximatable number as its members), data partitioning may degrade the application's performance because the data structure must be split to separate memory regions that have different error rates. This paper is the first to conduct an analysis of realistic C/C++ code to assess the impact of this problem. First, we find the type of data (e.g., "int", "struct point") that is assessed by the instruction that incurs the largest number of cache misses in a benchmark, which we refer to as the target data type. Second, we qualitatively estimate if the target data type of an application has approximate data and critical data interleaved. To this end, we set up three criteria to analyze it because definitively distinguishing a piece of data as approximate data or critical data is infeasible since it depends on each use-case. We analyze 11 memory intensive benchmarks from SPEC CPU 2006 and 2 graph analytics frameworks, and show that the target data types of 9 benchmarks are either a C struct or a C++ class (criterion 1). Among them, two have a pointer and a non-pointer member together (criterion 2) and three have a floating point number and other members together (criterion 3).